Data-Driven Finite Elements for Geometry and Material Design
Desai Chen1
1
MIT CSAIL
David I.W. Levin12
2
Disney Research
Shinjiro Sueda123
3
Wojciech Matusik1
California Polytechnic State University
Figure 1: Examples produced by our data-driven finite element method. Left: A bar with heterogeneous material arrangement is simulated
15x faster than its high-resolution counterpart. Left-Center: Our fast coarsening algorithm dramatically accelerates designing this shoe sole
(up to 43x). Right-Center: A comparison to 3D printed results. Right: We repair a flimsy bridge by adding a supporting arch (8.1x) speed-up.
We show a High-Res Simulation for comparison.
Abstract
1
Crafting the behavior of a deformable object is difficult—whether it
is a biomechanically accurate character model or a new multimaterial 3D printable design. Getting it right requires constant iteration,
performed either manually or driven by an automated system. Unfortunately, previous algorithms for accelerating three-dimensional
finite element analysis of elastic objects suffer from expensive precomputation stages that rely on a priori knowledge of the object’s
geometry and material composition. In this paper we introduce
Data-Driven Finite Elements as a solution to this problem. Given a
material palette, our method constructs a metamaterial library which
is reusable for subsequent simulations, regardless of object geometry
and/or material composition. At runtime, we perform fast coarsening of a simulation mesh using a simple table lookup to select the
appropriate metamaterial model for the coarsened elements. When
the object’s material distribution or geometry changes, we do not
need to update the metamaterial library—we simply need to update
the metamaterial assignments to the coarsened elements. An important advantage of our approach is that it is applicable to non-linear
material models. This is important for designing objects that undergo finite deformation (such as those produced by multimaterial
3D printing). Our method yields speed gains of up to two orders of
magnitude while maintaining good accuracy. We demonstrate the
effectiveness of the method on both virtual and 3D printed examples
in order to show its utility as a tool for deformable object design.
Objects with high-resolution, heterogeneous material properties are
everywhere: from the output of multimaterial 3D printers to virtual
characters gracing the screen in our summer blockbusters. Designing such objects is made possible by the tight coupling of design
tools and numerical simulation which allows designers (or automatic
algorithms) to update geometry or material parameters and subsequently estimate the physical effects of the change. Fast, accurate
simulation techniques that can handle runtime changes in geometry
and material composition are a necessity for such iterative design
algorithms.
CR Categories: I.3.7 [Computer Graphics]: Three-Dimensional
Graphics and Realism—Animation
Keywords: Data-driven simulation, finite element methods, numerical coarsening, material design
Introduction
The gold standard technique for estimating the mechanical behavior
of a deformable object under load is the finite element method (FEM).
While FEM is accurate, its solution process is notoriously slow,
making it a major bottleneck in the iterative design process. For this
reason, there have been a large number of works on speeding up FEM
simulations, and these speed improvements have enabled FEM to be
used in many performance critical tasks such as computer animation,
surgical training, and virtual/augmented reality. Unfortunately, even
though techniques such as model reduction or numerical coarsening
can achieve order-of-magnitude performance increases, they require
expensive precomputation phases, typically on the order of minutes
for large meshes. This precomputation requires knowledge of an
object’s geometry and material composition a priori, something
that is not known during a design task. When the user updates the
model by changing the geometry or the material distribution, the
preprocessing step must be run again. As shown in Fig. 2a, since
this step is inside the design loop, the user cannot get rapid feedback
on the changes made to the object.
We propose Data-Driven FEM (DDFEM), a new simulation methodology that removes these limitations and is thus extremely wellsuited to the types of design problems discussed above. We divide
an object into a set of deformable voxels using embedded finite
elements and coarsen these voxels hierarchically. A custom metamaterial database is populated with materials that minimize the error
incurred by coarsening. This database is learned once in a completely offline fashion and depends only on the set of materials to be
used by the deformable object and not on the actual material distribution and geometry. At runtime we use the database to perform fast
coarsening of an FEM mesh in a way that is agnostic to changes in
geometry and material composition of the object. The key features
of the algorithm are its ability to handle arbitrary, nonlinear elastic
Offline Preprocessing
High-resolution Mesh
Typical
Methods
Online simulation
Reduced Mesh
Simulated
Design Changes
(a) Typical methods
Online Coarsening
Offline Database
Construction
High-resolution Mesh
Our Method
Online Simulation
Reduced Mesh
Simulated
Design Changes
(b) Our method
Figure 2: (a) In a typical method, the preprocessing step is offline,
making the design loop slow. (b) In our method, we move the timeconsuming offline computation outside of the design loop.
constitutive models as well as to avoid expensive precomputation
within the design loop (Fig. 2b). DDFEM is the first algorithm
optimized for interactive geometric and material design problems.
2
Related Work
Efficient FEM simulation plays an important role in designing deformable objects. As mentioned, these problems are common in
engineering and graphics [Bendsoe and Sigmund 2003; Bickel et al.
2010; Kou et al. 2012; Skouras et al. 2013; Chen et al. 2013; Xu
et al. 2014]. We can broadly partition the space of approaches for
optimizing FEM simulation into two categories. We term the first
category numerical approaches. These methods use fast matrix inversion techniques and other insights about the algebra of the finite
element method to increase its performance. Simulators based on
the multigrid method [Peraire et al. 1992; Zhu et al. 2010; McAdams
et al. 2011] and Krylov subspace techniques [Patterson et al. 2012]
have yielded impressive performance increases. Other hierarchical
numerical approaches, as well as highly parallel techniques, have
also been applied to improve the time required to perform complex
simulations [Farhat and Roux 1991; Mandel 1993]. Finally, Bouaziz
et al. [2014] propose specially designed energy functions and an
alternating time-integrator for efficient simulation of dynamics.
The second set of methods are reduction approaches. These algorithms attempt to intelligently decouple or remove degrees of
freedom (DOFs) from the simulated system. This leads to smaller
systems resulting in a massive increase in performance, with some
decrease in accuracy. Our algorithm falls into this category. Note,
however, that numerical and reduction approaches need not be mutually exclusive. For example, our algorithm may potentially be used
as a preconditioner for a numerical approach. Algorithms based on
reduction approaches mitigate the inevitable increase in error using
one or more of three approaches: Adaptive remeshing, higher-order
shape functions, or by adapting the constitutive model.
Adaptive remeshing alters the resolution of the simulation discretization in response to various metrics (stress, strain etc.). Such methods
seek to maintain an optimal number of elements and thus achieve
reasonable performance. Adaptive remeshing has proven useful
for simulating thin sheets such as cloth [Narain et al. 2012], paper [Narain et al. 2013], as well as elastoplastic solids [Wicke et al.
2010] and solid-fluid mixtures [Clausen et al. 2013]. More general
basis refinement approaches have also been suggested [Debunne
et al. 2001; Grinspun et al. 2002]. While these methods do improve
the performance of simulation algorithms, they have some drawbacks. First, they often require complicated geometric operations
which can be time consuming to implement. Second, they introduce
elements of varying size into the FEM discretization. This can lead
to poor numerical conditioning if not done carefully. Finally, in order to maintain accuracy, it may still be necessary to introduce many
fine elements, leading to slow performance. Alternatively, one can
turn to P-Adaptivity for help. This refers to adaptively introducing
higher-order basis functions in order to increase accuracy during
simulation [Szabó et al. 2004]. Unfortunately, these methods suffer
from requiring complicated mesh generation schemes and are not
well-suited for iterative design problems.
An alternative approach to remeshing is to use higher order shape
functions in order to more accurately represent the object’s motion
using a small set of DOFs. Modal simulation techniques fall into
this category [Shabana 1991; Krysl et al. 2001; Barbič and James
2005]. Substructuring [Barbič and Zhao 2011] decomposes an input
geometry into a collection of basis parts, performing modal reduction
on each one. These basis parts can be reused to construct new
global structures. Other approaches involve computing physically
meaningful shape functions as an offline preprocessing step. For
instance, Nesme et al. [2009] compute shape functions based on
the static configuration of a high resolution element mesh induced
via a small deformation of each vertex. Faure et al. [2011] use
skinning transformations as shape functions to simulate complex
objects using a small number of frames. Gilles et al. [2011] show
how to compute material aware shape functions for these framebased models, taking into account the linearized object compliance.
Both Nesme et al. [2009] and Faure et al. [2011] accurately capture
material behavior in the linear regime, but, because their shape
functions cannot change with the deformed state of the material,
they do not accurately capture the full, non-linear behavior of an
elastic object. Our non-linear metamaterials rectify this problem.
Computing material aware shape functions improves both the speed
and accuracy of the simulation. However, these methods require a
precomputation step that assumes a fixed material distribution and
geometry. If the material distribution changes, these shape functions
must be recomputed, and this becomes a bottleneck in applications
that require constantly changing material parameters.
The final coarsening technique involves reducing the degrees of freedom of a mesh while simultaneously augmenting the constitutive
model at each element, rather than the shape functions. Numerical
coarsening is an extension of analytical homogenization which seeks
to compute optimal, averaged material for heterogeneous structures
[Guedes and Kikuchi 1990; Farmer 2002]. Numerical coarsening,
for instance, has been applied to linearly (in terms of material displacement) elastic tetrahedral finite elements [Kharevych et al. 2009].
These methods require an expensive precomputation step (a series
of static solves) that must be repeated when the material content, or
the geometry of an object changes. This holds these methods back
from being suitable for iterative design problems.
Recently, three methods have been introduced that are similar to
ours in spirit. Bickel et al. [2009] measured force-displacement to
compute a spatially varying set of Young’s moduli, interpolated in
strain space. Our work also involves learning new constitutive models for finite element methods with several key differences. First, we
present a more robust energy-based metamaterial model that does not
require incremental loading during simulation. Second, the previous
work relies on captured data to build constitutive model, while we
use a new sampling strategy that allows us to build our metamaterial
model virtually. This allows us to leverage large, high-performance
compute clusters to speed up the process. Finally, their work is
completely geometry dependent—their computed material models
cannot be transferred to new meshes. Xu et al. [2014] and Li et
al. [2014] computed material distribution given user specified forces
Offline Database Construction
Online Simulation
Metamaterial Fitting
(Section 4.3)
Fast Lookup
Compression
(Section 4.4)
Runtime Coarsening
(Section 5)
Simulation
Figure 3: An overview of DDFEM. The method is divided into an offline database construction phase and an online simulation phase. During
database construction we build a collection of metamaterials which can be used for accurate coarsening. These metamaterials are compressed
(optional) and stored in a database, indexed by material ID. During online simulation we coarsen a high-resolution simulation mesh by
performing a fast lookup into the precomputed database. Finally, finite element simulation is performed on this coarsened mesh.
and displacements. They compute materials in a low-dimensional
space of material modes to speedup and regularize the solution.
However, their method requires a known geometry, and furthermore,
users cannot control per-element material assignment. Rather than
computing material distribution, our method supports user specified topology changes and material assignment. These features are
important due to the design-centric nature of our work.
Data-driven techniques have also been applied to add fine detail to
coarse surgical simulations [Cotin et al. 1999; Kavan et al. 2011;
Seiler et al. 2012]. While these methods do not help with the type of
iterative design problems addressed here, they could be combined
with our fast coarsening in order to quickly “upscale” coarsened
simulation results to original simulation resolution.
2.1
Contributions
This work makes the following technical contributions:
• The first algorithm for fast runtime coarsening of arbitrary,
nonlinear, elastic material models for finite element analysis.
• A compact metamaterial model used for coarsening.
• An efficient procedure for fitting our metamaterial model.
3
Overview
DDFEM is a combination of embedded finite elements and hierarchical coarsening. In this section, we discuss the problem of coarsening
and introduce the notion of a material palette. We conclude by
summarizing the two main stages of DDFEM—offline metamaterial
construction and online coarsening.
The key component of our
DDFEM is coarsening. Coarsening involves reducing the number
of vertices in a finite element simulation mesh in order to improve
runtime performance. Since simply removing vertices can greatly reduce the accuracy of the simulation, coarsening schemes also assign
new materials to coarsened elements to minimize this effect.
Coarsening for finite elements
We regard the global coarsening of a simulation mesh as the result
of many local coarsening operations which map from contiguous
subsets of fine elements with applied materials to coarse elements
with new, optimized materials. Our goal is to precompute these
optimized materials so that coarsening is fast at runtime. Below we
discuss how to make such a precomputation tractable beginning with
our choice of Finite Element simulation methodology.
The defining feature of conforming finite element methods is that the simulation
mesh is aligned with the geometry of the object being simulated.
Conforming vs. embedded finite elements
One obvious feature of conforming meshes is that the mesh itself is a
function of the input geometry. This means that the output of a local
coarsening operator (the coarsened mesh) will also be a function
of the input geometry. Also, the new material computed by each
local coarsening operator will be a function of input geometry. This
dependence on input geometry is a significant issue to overcome
if we wish to precompute coarsened materials because, in design
problems, the input geometry is in constant flux. The number of
precomputed coarse materials now depends on the local material
assignment on the simulation mesh and the input geometry. Thus
space of coarsened materials is prohibitively large. To mitigate this
we turn to embedded finite elements. These methods embed the
geometry to be simulated into the simulation mesh with no regard
for whether the mesh conforms to the geometry or not. Thus an
identical simulation mesh can be used for any input geometry. Local
coarsening operations on the embedded mesh yield identical coarse
elements and the optimized coarse material depends only on the
local material distribution on the simulation mesh. This significantly
reduces the size of the coarsened material space. In this paper we
embed all simulation geometry into a hexahedral simulation mesh.
We further shrink the space of coarsening operators using an observation about material design. Designers do
not work in a continuous space of materials but limit themselves to
a relatively compact set (e.g. rubber, wood, steel) related to their
problem domain (Fig. 4). We call these discrete sets of materials
palettes and denote them P = {M0 , M1 , . . . , Mn }. Here Mi denotes a specific material model in P, and n is the size of the material
palette. In this work we limit ourselves to (nonlinear) hyper-elastic
materials, which means that each Mi can be represented by a strain
energy density function. We also include a void (or empty) material
in every palette. This allows us to perform topology changes in the
same manner in which we perform material assignment updates. In
subsequent sections, we use a left superscript to indicate the level of
coarsening. For example, 0 P denotes a material palette at the fine
scale while 1 P denotes the new palette of metamaterials that results
from the first coarsening step.
Material palette
Stratysys Objet Connex 500 Palette
Rubber
ABS
Rigid
Empty
Figure 4: An example material palette for the Stratasys Objet Connex 500 3D printer.
With the material palette in hand, we can now define
our algorithm, which is divided into two distinct phases: an offline database construction stage and an online coarsening stage.
Below we detail the input, output, and steps of each stage:
Algorithms
Online Coarsening
• INPUT: High resolution hexahedral simulation mesh with
material IDs and
coarsened hexahedral simulation mesh
• OUTPUT: Metamaterial assignments for coarse mesh
• STEPS:
• FOR EACH 2×2×2 block in the high resolution mesh
• Replace with single coarse element
• Assign material from 1 P using high resolution
material IDs as database key
• END
We stress that both stages of the
DDFEM algorithm can be applied hierarchically. Given the first
level of metamaterials, 1 P, we can construct a metamaterial library,
2
P, for the second level by using 1 P as an input material palette.
At runtime, the coarsening algorithm looks up materials from 2 P to
replace each 2×2×2 coarse block with a single element.
Hierarchical coarsening
Having introduced the broad strokes of the DDFEM scheme, we
move on to a detailed explanation of each algorithmic component.
First we discuss database construction in §4, followed by the runtime
component in §5. We end by demonstrating the speed and accuracy
of DDFEM in §6.
Metamaterial Database Construction
We construct our metamaterial database using a potential energy
fitting approach. This is valid due to the hyperelastic materials that
make up our material palettes. Material fitting considers 2×2×2
blocks of high-resolution hexahedral elements (denoted 0 E). For
each element 0 Ek ∈ 0 E, its material is referred to as 0 Mk ∈
0
P. (Note that E refers to a set of elements and E refers to a
single element.) Given 0 E, we can sample its deformation space,
and using 0 Mk , compute the potential energy 0 V for each sample.
Now we must find a metamaterial that, when applied to a single
coarse element 1 E best approximates 0 V . This is accomplished by
fitting a metamaterial potential energy function , 1 V , to the set of
deformation/energy samples. The fitted metamaterial is stored in the
metamaterial database and indexed by the material indices of 0 M.
4.1
Metamaterial Model
Our fitting approach depends on choosing a good metamaterial
model. In order to ensure that our model meets the criteria for a
material energy function [Marsden and Hughes 2012], we choose
our metamaterial model for 1 E as a combination of material models
V3
0
V2
Coarsen
V4
X1
1
X3
1
1
1
High-Resolution Elements 0Ɛ
X2
X4
Coarse Element 1E
Figure 5: The relationship between high-resolution and coarsened elements. At each quadrature point 1 X k , the coarse element
copies the corresponding energy density function 0 Vk from the highresolution element.
of 0 Mk :
1
V (1 u, 1 p) =
8
X
wk 0 Vk (0 uk , 1 pk , 1 X k ),
(1)
k=1
where 0 Vk is the strain energy density of 0 Mk at quadrature point
position 1 X k (Fig. 5). Here 1 u is the vector of nodal displacements associated with 1 E while 0 uk are displacements for the kth
element at level 0 reconstructed using trilinear interpolation from
1
u. The vector 1 p stores the material parameters for the coarse
metamaterial and consists of the stacked material parameter vectors for each material in 0 Mk , themselves denoted by 1 pk . wk is
the standard Gaussian quadrature weight. We note that our model
incurs slight computational overhead at runtime because we must
evaluate potential energy functions at 8 quadrature points. However,
the speed improvement gained by coarsening makes the remaining,
per-quadrature point expense negligible.
We observe that even if the individual base material models are
isotropic, the metamaterial can become anisotropic by assigning
different material parameters at the quadrature points. We counter
this by augmenting the metamaterial model with an anisotropic term,
which improves fitting. The complete model is then given by
8 X
1
V (1 u, 1 p, C) =
wk 0 Vk (0 u, 1 pk , 1 X k )
k=1
(2)
q
2 T
T
+ Ck
v F k F kv − 1
,
where v is a unit-length direction of anisotropy and Ck is the scaling
parameter at the kth quadrature point.
4.2
4
0
0
Offline Database Construction
• INPUT: A palette of materials to be applied to high-resolution
hexahedral simulation meshes 0 P
• OUTPUT: A new palette of coarse metamaterials, 1 P, and
a mapping from fine material combinations to the coarsened
materials in 1 P.
• STEPS:
• FOR EACH material combination applied to a 2×2×2 cube
of high resolution elements
• Sample potential energy function of 2×2×2 block
• Fit metamaterial for coarse hexahedral element
• Add metamaterial to 1 P using high resolution
material IDs as database key
• END
V1
0
Force Space Sampling
As mentioned previously, we take a sampling-based approach to
metamaterial fitting. In order to fit our model (Eq. 2) to 0 V we
first draw a number of samples from the deformation space of 0 E
and compute 0 V for each sample. If a user has prior knowledge
of the set of meshes and simulations that they will require, then
the best way to draw the samples is to run a number of anticipated
simulations with various material combinations. In this paper, we
provide a more general method to draw samples for a metamaterial.
Initially, we attempted sampling by applying a random deformation
to the corners of 1 E ; however, this led to many infeasible samples
for very stiff materials. In order to alleviate this problem we perform
sampling in the force space.
For each element 0 E ∈ 0 E we apply a set of randomly generated
forces. We solve an elastostatic problem to compute the deformation
of 0 E, using constraints to remove rigid motion. Recall that this is
0
fast because
of just 8 elements. Each sample is then
0 0E consists
a tuple u, V (Fig. 6) where 0 u are the nodal displacements
of 0 E, and 0 V is the strain energy density value of this deformed
configuration.
Force Space
Energy
SIMULATE
(1) Choose
Random
Material
(2) Choose
Furthest
Material
Deformation
Energy
Energy
Figure 6: We sample the energy function of a 2×2×2 block of
hexahedra by applying random forces and deforming the block using
FEM. Each set of forces results in a deformation-energy tuple which
is used during fitting.
FIT
Deformation
Deformation
Figure 7: Metamaterial potential energy functions are fitted to the
deformation-energy samples using a least-squares optimization.
4.3
Fitting
Given a set of deformation samples, 0 u, 0 V , we perform a least
squares fit to determine the parameters, 1 p, for a given metamaterial
(Fig. 7):
ns
X
2
1 ∗
0
p = argmin
Vs − 1 V r 0 us , 1 p
,
(3)
1p
s=1
where r constructs 1 u from 0 u, ns is the total number of samples,
and s indexes all samples. In our experiments we use the simplest
form of r choosing it to extract the displacements of the corners of
0
E.
If performed naively this
optimization is nonlinear because we must simultaneously solve
for v k , the preferred direction of anisotropy. This can severely
slow the fitting procedure, especially in cases where it would otherwise be a linear least squares problem (i.e if all fine-scale materials
are Neo-Hookean or a similarly simple material model). To avoid
this problem we first estimate all anisotropy directions, and then
solve Eq. 3 for the remaining material parameters. Our intuition
is that anisotropy manifests itself as preferential stretching along
a particular axis. To find this axis, we apply stretching forces to a
block in a discrete set of directions uniformly sampled over a sphere.
If the stretching force is close to the direction of anisotropy, then
the amount of stretching deformation is reduced. For any given
stretching direction v, we apply a stretching force and compute the
deformation gradient F of each quadrature point. Under F, a unit
length vector in direction v is stretched to a new length l = kFvk.
The set of all 3D vectors lv forms an ellipse-like shape. We find the
principal axes of the ellipse (via SVD) and use them as directions of
anisotropy.
Fitting in the Presence of Anisotropy
Since vastly different material assignments, 0 Mk ,
can produce the same metamaterial, our naı̈ve cost function (Eq. 3)
can produce very large parameter values and even non-physical
negative ones. For example, consider a homogeneous material assignment at the high-resolution level. The same metamaterial can be
achieved by interleaving hard and soft materials at each fine element
or by assigning a single, well chosen material to all fine elements. To
Regularization
Compressed
Database
Uncompressed
Database
(3) Repeat
Figure 8: Database compression step adds metamaterials to a
compressed database by a farthest point sampling strategy.
overcome this, we add a regularization term to control the parameter
ranges and prevent overfitting of the training samples. Our modified
error function takes the following form:
ns
X1
X
2
0
Vs − 1 V r 0 us , 1 p
+λ
( pk − 0 pk )2 , (4)
s=1
k
which prevents material parameters from deviating too far from
0
Mk . We chose the regularization constant λ = 0.01 for the results
in this paper. In our experiments, since the base energy functions are
linear with respect to the material parameters, the fitting problem
can be solved by linear regression with regularization.
4.4
Database Compression
Given n materials in the palette, the number of material combinations in a 2×2×2 block is n8 . In modern hardware, it is impossible to compute and store all material combinations even for a
moderately-sized palette with 100 materials. In order to compress
the number of materials stored in our metamaterial database, we
select a small number of representative material combinations and
remove all others. We compare materials in a feature space. In order
to construct metamaterial feature vectors, we first select a common
subset of deformations from all computed deformation samples. We
then evaluate the potential energies of each metamaterial at each
deformation sample. The stacked vector of energies becomes a
metamaterial feature vector.
Since our base materials differ in stiffness by orders of magnitude,
we take the logarithm to measure the difference in ratio. Let D be
the L2 norm of log-energies between the two materials given by
sX
D(A, B) =
(log(0 VsA ) − log(0 VsB ))2 ,
(5)
s
where A and B denote two distinct metamaterials in the database.
Given the distance metric, we can select k representatives materials using farthest point sampling [Eldar et al. 1997]. We randomly
choose an initial metamaterial and then repeatedly select the material
combination furthest away from any previous representatives – continuing until we obtain k representatives (Fig. 8). This compression
algorithm chooses k samples that equally cover the metamaterial
energy space, helping to preserve good behavior in our coarse simulations.
4.5
Hierarchical Coarsening
While one level of coarsening can yield significant speed-ups,
DDFEM can also be applied hierarchically. As discussed in §4.4, the
exponential growth of metamaterials palettes at each level makes it
prohibitively expensive to perform fitting. We address this by changing our coarsening strategy. Instead of choosing 0 E to be a 2×2×2
block we choose it to be a 2×1×1 block, which we coarsen. We construct an intermediate database of materials and compress. We then
choose 0 E to be a 1×2×1 block, coarsen and compress, and finally
a 1×1×2 block, coarsen and compress. Intermediate compression
greatly reduces the number of samples we need to generate in order
to populate the material parameter database for the next coarsening
level. It is important to note that our intermediate databases only
store lookup tables which allow us to extract appropriate material
IDs for the next coarsening stage. Material parameters need only
be stored in the final database since it is these elements that are
simulated.
Level i
Level ix
Coarsen x
Database
for level i
Level i+1
Coarsen y
Database
for level ix
Database
for level i+1
Figure 9: Hierarchical coarsening operates on one dimension at a
time, performing clustering at each intermediate stage (here denoted
ix ). This allows our compression algorithm to be applied aggressively, greatly reducing the number of energy samples we need for
fitting material parameters.
5
Runtime Simulation
Once our metamaterial database, 1 P, has been constructed we can
use it to perform fast online coarsening. Initially, the user loads
geometry which is embedded in a hexahedral grid for simulation.
Prior to simulation we iterate over all 2×2×2 blocks of hexahedral
elements and perform mesh coarsening by replacing these 8 elements with a single coarse element. We perform a database lookup
into 1 P, using the material ID numbers of the 8 original elements,
to quickly retrieve the optimal metamaterial for this coarse element.
Database lookup is fast (even using our unoptimized, single-threaded
implementation), and this is what makes DDFEM so appealing. We
achieve significant simulation speed-up from coarsening, retain accuracy in the simulation, and reduce the cost of material coarsening
at runtime to a negligible amount. Our material model can be used
in any simulation algorithm suitable for non-linear elasticity. In
our experiments, we use Coin-IpOpt [Wächter and Biegler 2006]
to implement static and dynamics simulations with tolerance (“tol”
option) set to 0.5. We use Pardiso as our linear solver. For timing
purposes, we limit Pardiso to single thread mode. The pseudo-code
for static simulation is shown in Alg. 1.
6
Results and Discussion
All the results shown here are simulated using nonlinear constitutive models at the fine scale. This and coarsening speed are the
key differentiating factors between DDFEM and other coarsening
algorithms such as Nesme et al. [2009] and Kharevych et al. [2009].
Our database starts with three Neo-hookean base materials with
Young’s modulus 1e5, 1e6, 1e7 and Poisson’s ration 0.45. For comparison with 3D-printed objects, we used two base materials with
measured Young’s moduli. We use 500 force directions, and sample
5 magnitudes in each direction, resulting in 2500 force samples for
each material combination. In addition, we generate 500 stretching
samples for computing the direction of anisotropy. During fitting,
we use shear modulus and Lamé’s first parameter, as well as the
spring stiffness. We repeat the same process for the second level of
coarsening, using 6561 materials in the first level as base materials.
Algorithm 1: Static Simulation
1: repeat
2:
f : global force vector
3:
L: triplet list for global stiffness matrix
4:
for each element e do
5:
compute elastic force fe
6:
add fe to f
7:
end for
8:
add external force fext to f
9:
for each element e do
10:
Ke : element stiffness matrix
11:
for each quadrature point q do
12:
compute stiffness matrix Kq at quadrature point
13:
Ke + = Kq
14:
end for
15:
append entries of Ke to L
16:
end for
17:
sort L to get sparse stiffness matrix K
18:
set entries in K and f for fixed vertices
19:
∆x = K−1 f
20:
compute step size h using line-search
21:
x+ = h∆x
22: until convergence
Example
Pushing(0)
Pushing(1)
Pushing(2)
Bending(0)
Bending(1)
Bending(2)
Twisting(0)
Twisting(1)
Twisting(2)
Buckling(0)
Buckling(1)
Buckling(2)
Fibers(0)
Fibers(1)
Fibers(2)
Bridge(0)
Bridge(1)
Bridge-arch(0)
Bridge-arch(1)
George(0)
George(1)
George-bone(0)
George-bone(1)
grid size
16×16×16
8×8×8
4×4×4
8×32×8
4×16×4
2×8×2
8×32×8
4×16×4
2×8×2
128×8×16
64×4×8
32×2×4
32×100×32
16×50×16
8×25×8
56177
9727
65684
11695
46152
6755
46152
6755
rel sp
1.0
11.5
31.4
1.0
12.6
22.7
1.0
15.2
20.7
1.0
50.1
331.8
1.0
51.2
489.5
1.0
8.4
1.0
8.1
1.0
16.4
1.0
13.2
time/iter
1.010
0.087
0.032
0.270
0.028
0.015
0.300
0.031
0.019
8.85
0.28
0.12
193.85
4.95
0.96
43.44
4.88
54.99
7.84
52.19
3.49
41.35
2.70
iters
5
5
5
28
22
22
16
10
12
32
20
7
17
13
7
14
15
3
3
23
21
12
14
error
8.91e-4
1.36e-2
5.60e-2
8.88e-2
1.56e-2
3.28e-2
7.24e-3
3.14e-2
2.94e-2
4.26e-2
4.39e-3
3.68e-4
2.86e-2
2.99e-2
Table 1: Relative performance, absolute performance in seconds
and average vertex error relative to the bounding box size for fullresolution and coarsened simulations. Relative performance illustrates the performance increase gained by coarsening with respect to
the time taken for the high-resolution static simulation to converge.
Bracketed numbers after each example name indicate the number of
coarsening levels with 0 indicating the high-resolution simulation.
All computation times are recorded using Coin-IpOpt running in
single threaded mode on a 2.5 GHz Intel Core i7 processor.
We select 400 representatives at each intermediate level.
6.1
Database
One advantage of our compact metamaterial representation is the
small amount of storage it requires. In fact we require only 6 × 8 =
48 floating-point values for each material at the first coarsening
level and 6 × 64 = 384 values for the second level. (For each
Pushing
12x Faster
Twisting
15x Faster
(a)
(b)
(c)
(d)
Figure 10: Examples of pushing a cube (a - Initial State, c - Compressed) and twisting a bar (b - Initial State, d - Compressed), both
with heterogeneous material distribution. We compare DDFEM to
Naı̈ve Coarsening and the ground-truth, High-Res Simulation. We
render wire frames to show the simulation meshes.
1 Level of Coarsening
13x Faster
(a)
(c)
2 Levels of Coarsening
23x Faster
1 Level of Coarsening
50x Faster
(a)
(b)
(c)
(d)
Figure 12: Compressing a heterogeneous slab using Naı̈ve Coarsening (1 level and 2 levels of coarsening), DDFEM (1 level and 2
levels of coarsening) and a High-Res Simulation. The top, darker
layer is stiffer, causing the object to buckle. The bottom vertices
are constrained to stay on the floor. Figure (a,b) shows the slabs
before compression, figure (c,d) shows the slabs after compression
and figure. Notice that, after 1 level of coarsening, Naı̈ve Coarsening neither compresses nor buckles as much as either DDFEM or
High-Res Simulation. After 2 levels of coarsening, the buckling behavior is lost. The Naı̈ve Coarsening fails to capture the compressive
behavior of High-Res Simulation, whereas DDFEM does.
1 Level of Coarsening
51x Faster
2 Levels of Coarsening
489x Faster
(a)
(b)
(c)
(d)
(b)
(d)
Figure 11: Bending a heterogeneous bar: We compare DDFEM
to Naı̈ve Coarsening and a High-Res Simulation. Subfigures (a, c)
show comparison for 1 level of coarsening, and (b, d) show 2 levels
of coarsening. The naı̈ve coarsening approach results in a much
stiffer behavior, whereas our fitted model more closely approximates
the fine model.
finer element, 0 p contains 2 material moduli plus C, v.) For further
recursive levels, we can limit ourselves to 320 values per material.
Our current 3 material database is 4 megabytes in size.
6.2
2 Levels of Coarsening
332x Faster
Simulation Results
We show results from elastostatic simulations performed using
DDFEM. We also demonstrate its performance advantages over
high-resolution simulations. We render wire frames to show the
discretizations of the high-resolution and coarse meshes. We first
show examples of two simple simulations, the pushing and twisting of a rectangular object with heterogeneous, layered material
distribution (Fig. 10). Note that in all cases DDFEM qualitatively
matches the behavior of the high-resolution simulation. We also
compare the performance of DDFEM to a naı̈ve coarsening method
that uses the material properties from 2×2×2 element blocks of the
high-resolution simulation mesh at each corresponding quadrature
point. In our supplemental video we compare to a second baseline
model which averages material parameters inside each coarse element. This average model is less accurate than the Naı̈ve model in
Figure 13: Simulating a bar with an embedded set of fibers using
Naı̈ve Coarsening, DDFEM and a High-Res Simulation. Note that
DDFEM captures the characteristic twisting motion of the bar better
than Naı̈ve Coarsening. (a,b) shows the initial state of both bars
while (c,d) shows the deformed state after pulling on the top of the
bars.
all cases.
Naı̈ve approaches often exhibit pathological stiffness for heterogeneous materials (illustrated by the lack of compression of the box
and lack of twisting of the bar) [Nesme et al. 2009]. In these cases,
DDFEM yields good speed ups while maintaining accuracy. For a
single level of coarsening we achieve 8 times or greater speed ups
for all examples. Performance numbers and mean errors are listed
in Table 1. Since the fine simulation and the coarse simulation have
different numbers of vertices, we create a fine mesh from the coarse
simulation by trilinearly interpolating the fine vertices using the
coarse displacements. The errors are measured by computing the average vertex distance relative to the longest dimension of the bounding box in rest shape. We also examine the behavior of DDFEM
8.1x
Squared Norm of Residual
8.4x
High−Res Simulation
1 Level of Coarsening
2 Levels of Coarsening
−2
10
Figure 16: Accelerating geometry change: We repair a structurally
unsound bridge by adding a supporting arch (8x faster).
−4
10
−6
10
0
1000
2000
Number of Iterations
3000
Figure 14: Comparison of CG iterations on high-resolution and
coarsened meshes of the George-bone example. The squared residual is measured as kKx − f k2 . We observe that CG converges much
faster on coarse meshes.
Example
Bending(0)
Bending(1)
Bending(2)
Buckling(0)
Buckling(1)
Buckling(2)
Fibers(0)
Fibers(1)
Fibers(2)
Add Support
and Coarsen
grid size
8×32×8
4×16×4
2×8×2
128×8×16
64×4×8
32×2×4
32×100×32
16×50×16
8×25×8
quad/iter
(12)
0.11
0.015
0.014
1.0
0.11
0.10
10.0
1.0
0.68
time/iter
(2-21)
0.27
0.028
0.015
8.8
0.28
0.12
193
4.9
0.96
iters
28
22
22
32
20
7
17
13
7
Table 2: Portion of time in seconds used by quadrature computation
during static simulation in seconds. Bracketed numbers indicate
corresponding lines in Algorithm 1.
during bending (Fig. 11). Yet again the naı̈ve coarsening method
completely fails to capture the behavior of the high-resolution result,
while DDFEM offers a much better approximation.
Table 2 shows time spent in quadrature
evaluation versus in solver at coarse and fine levels. We use 8
quadrature points for the first level of coarsening and 64 for the
second level. The time for computing the local stiffness matrix for
one element increases from 0.1ms to 1ms. In the second level, the
speedup comes from the reduced number of elements over which to
perform quadrature and the time required for the linear solver.
Performance Analysis
To further investigate the performance of our coarsened simulations
we replaced Pardiso with an assembly-free Jacobi-preconditioned
conjugate gradient (CG) linear solver and used this to simulate
our George-bone test case. While the overall runtime of the highresolution simulation increased from 496 to 2082 seconds (most
likely do to the unoptimized nature of our solver) our coarsened
model achieved 20x and 67x speedups using one level and two
levels of coarsening respectively. One might expect no benefit from
the second level of coarsening since the number of quadrature points
remains constant. However, the number of CG iterations is roughly
proportional to the number of vertices in the simulation mesh and
thus the coarse model converges more quickly (Fig. 14). Since our
metamaterial models are not restricted to use a fixed number of
quadrature points, one could design coarse models that are more
tailored towards assembly-free solvers by reducing the number of
quadrature points and simplifying the strain energy expressions.
DDFEM can capture the gross behavior of complex, spatially-varying material distributions. Fig. 12
shows the results of applying DDFEM to a non-linearly elastic slab
Complex material behavior
with a stiff “skin.” The bottom of the slab is constrained to slide
along the ground with one end fixed. When force is applied to
the free end of the slab, buckling occurs. DDFEM captures the
gross behavior of the bar and approximates the overall amount of
compression well. However, it cannot replicate the frequency of
the high-resolution buckling pattern due to the coarseness of the
simulation mesh. Fig. 12 also shows a comparison with 2nd level
coarsening. In this case, the overall compression of the bar is still
captured accurately. For this example DDFEM affords 50 times
(1 level of coarsening) and 332 times (2 levels of coarsening) performance improvements over the high-resolution simulation. The
artificial stiffness of the naı̈ve model can be seen in the reduced buckling and compression when compared to DDFEM at both coarsening
levels.
Next we explore the ability of
DDFEM to handle highly anisotropic material distributions (Fig. 13).
We embed a helical set of stiff fibers in a soft, non-linearly elastic
matrix. Pulling on the object induces a twisting. Again, at one coarsening level DDFEM captures this anisotropic behavior well, much
better than the naive approach, and gains a 51 times speed up over
the high-resolution simulation. Worth noting is that the DDFEM
bar is slightly softer in the y-direction. This kind of inaccuracy
should be expected. Since our method builds a low-dimensional
approximation of a potential energy function we cannot hope to
accurately reproduce the complete behavior of the high-resolution
simulation. What is important is that DDFEM captures the salient
global behavior, in this case, the twisting of the bar.
Anisotropic material distribution
We present three examples of
using DDFEM for geometry and material design. In the first example, we edit the material composition of the sole of a running
shoe in order to stabilize it. Fig. 15 shows the effect of the three
material edits as well as relative speed up achieved over the full
resolution simulation and coarsening time. DDFEM performance is
always an order of magnitude more than that of the high-resolution
simulation, and, most importantly, our coarsening times are on the
order of milliseconds. We stress that our current implementation is
completely single threaded and that coarsening, which in our case involves a simple database lookup, is inherently parallel. In the second
example, we add a supporting arch to a bridge. Prior to the addition
of the support structure, the bridge sags catastrophically. The fast
coarsening of DDFEM allows us to achieve an 8 fold increase in
simulation performance using a single coarsening pass. In the third
example, we add a rigid skeleton to a deformable character (George)
in order to control his pose. Here our single threaded, data-driven
coarsening only takes 200ms.
Geometry and material design
Though the examples shown in this paper focus on
static analysis, DDFEM is equally applicable to dynamic simulations.
At its core, DDFEM simply supplies new, more accurate material
models for use during simulation. This makes the method useful for
accelerating various animation tasks as well. In the accompanying
videos we show a dynamic simulation of our fiber embedded bar,
computed using a standard linearly-implicit time integrator.
Dynamics
12x
20x
34x
Coarsening
Coarsening
72ms
Original Material
43x
Coarsening
73ms
Horizontal Fibers
72ms
Vertical Fibers (Heel)
Vertical Fibers (Toe)
Figure 15: Designing a shoe sole: We compare the performance of DDFEM to that of High-Res Simulation in the context of a material design
problems. Large images show the effect of material changes on the sole of the shoe, which is being deformed under a “foot-like” pressure field.
Inset images show the materials assigned to the shoe sole and the embedded finite element simulation mesh. Numbers within arrows show
coarsening times between editing steps and the numbers in the upper left corner of each image show the relative performance of DDFEM to
High-Res Simulation. While the DDFEM sole is made up of many metamaterials, we display it as a single color to distinguish it from the
High-Res Simulation.
boundary representation to the method, in order to more correctly
handle non-axis aligned boundary conditions, could also be explored.
Third, the method acts on discrete materials. While we believe that
this is reasonable, considering the way that engineers and designers
approach material design, a method that coarsens continuous spaces
of non-linear materials could be beneficial.
Figure 17: Correcting George’s posture using a rigid skeleton
(High-Res Simulation and DDFEM).
Figure 18: A comparison of DDFEM (2 levels of coarsening) to real
world deformations of 3D printed, multi-material designs. DDFEM
captures the twisting behavior of an anisotropic bar much more
accurately than Naı̈ve Coarsening. Similarly DDFEM accurately
predicts the deformation of our heterogeneous George character.
6.3
Fabricated Results
Finally, we test the accuracy of our simulation against fabricated
results, created using a Stratysys Object Connex 500 multimaterial
3D printer. We fabricated a bar with embedded helical fibers as well
as our George character and applied specified loads to both. We
show qualitative comparison of the deformed configurations of these
real-world examples to our simulated results (2 levels of coarseningFig. 18). Note that the simulation does an excellent job of predicting
the deformed configuration of both objects.
7
Limitations and Future Work
Because DDFEM relies on a database compression step to combat
the combinatorial explosion of metamaterials, accuracy is heavily
influenced by the set of representative metamaterials. Finding a
better way to select metamaterial structures is an interesting area
of future work. Second, in our attempt to make our method geometry independent, some accuracy when dealing with partially
filled boundary finite elements is sacrificed. Adding a parameterized
Many avenues of future work are promising. First, one could explore
topologically aware meshing (e.g. Nesme et al. [2009]) to allow
better handling of models with large empty regions. In fact shape
function learning approaches, such as Nesme et al. [2009] could
be combined with our material learning approach to produce even
more accurate simulations. Including these shape functions in our
database could, for instance, allow us to capture the wrinkles in our
buckling example. Second, extending DDFEM to more complex
material models, such as those involving plasticity, would be a useful
exercise. Third, DDFEM can be combined with an adaptive voxel
grid as well as other dynamic meshing approaches to obtain further
speed-ups. Finally, exploring hierarchical solvers based on DDFEM
coarsening is a very attractive direction. Solvers such as multigrid
methods rely on good coarse approximations to accelerate fine scale
simulations. Using DDFEM for these approximations could improve
the convergence rate, and thus the performance of such algorithms.
8
Conclusions
In this paper we presented the data-driven finite element method,
a sampling-based coarsening strategy for fast simulation of nonlinearly elastic solids with heterogeneous material distributions. Our
method is unique in that changing material parameters requires no
additional computation at runtime, making it well-suited for iterative
design problems. We have shown the utility of our method on
several examples wherein it garners a 8 to 489 times speed-up over
corresponding high-resolution simulations.
Acknowledgements
Thank you to the Anonymous Reviewers, Etienne Vouga, Maria
Shugrina and Melina Skouras for helpful suggestions and James
Thompkin for video assistance. This research was funded by NSF
grant CCF-1138967 and DARPA #N66001-12-1-4242.
References
BARBI Č , J., AND JAMES , D. L. 2005. Real-time subspace integration for St. Venant-Kirchhoff deformable models. ACM Trans.
Graph. 24, 3 (July), 982–990.
BARBI Č , J., AND Z HAO , Y. 2011. Real-time large-deformation
substructuring. ACM Trans. on Graphics (SIGGRAPH 2011) 30,
4, 91:1–91:7.
B ENDSOE , M., AND S IGMUND , O. 2003. Topology Optimization:
Theory, Methods and Applications. Engineering online library.
Springer.
B ICKEL , B., B ÄCHER , M., OTADUY, M. A., M ATUSIK , W., P FIS TER , H., AND G ROSS , M. 2009. Capture and modeling of
non-linear heterogeneous soft tissue. ACM Trans. Graph. 28, 3
(July), 89:1–89:9.
B ICKEL , B., B ÄCHER , M., OTADUY, M. A., L EE , H. R., P FIS TER , H., G ROSS , M., AND M ATUSIK , W. 2010. Design and
fabrication of materials with desired deformation behavior. ACM
Trans. Graph. 29, 4 (July), 63:1–63:10.
B OUAZIZ , S., M ARTIN , S., L IU , T., K AVAN , L., AND PAULY, M.
2014. Projective dynamics: Fusing constraint projections for fast
simulation. ACM Trans. Graph. 33, 4 (July).
C HEN , D., L EVIN , D. I. W., D IDYK , P., S ITTHI -A MORN , P., AND
M ATUSIK , W. 2013. Spec2Fab: A reducer-tuner model for
translating specifications to 3D prints. ACM Trans. Graph. 32, 4
(July), 135:1–135:10.
C LAUSEN , P., W ICKE , M., S HEWCHUK , J. R., AND O’ BRIEN ,
J. F. 2013. Simulating liquids and solid-liquid interactions with
Lagrangian meshes. ACM Trans. Graph. 32, 2 (Apr.), 17:1–17:15.
C OTIN , S., D ELINGETTE , H., AND AYACHE , N. 1999. Realtime elastic deformations of soft tissues for surgery simulation.
IEEE Transactions on Visualization and Computer Graphics 5, 1,
62–73.
D EBUNNE , G., D ESBRUN , M., C ANI , M.-P., AND BARR , A. H.
2001. Dynamic real-time deformations using space & time adaptive sampling. In Proc. SIGGRAPH ’01, Annual Conference
Series, 31–36.
E LDAR , Y., L INDENBAUM , M., P ORAT, M., AND Z EEVI , Y. Y.
1997. The farthest point strategy for progressive image sampling.
Image Processing, IEEE Transactions on 6, 9, 1305–1315.
FARHAT, C., AND ROUX , F.-X. 1991. A method of finite element
tearing and interconnecting and its parallel solution algorithm.
International Journal for Numerical Methods in Engineering 32,
6, 1205–1227.
ACM Trans. Graph. 28, 3 (July), 51:1–51:8.
KOU , X., PARKS , G., AND TAN , S. 2012. Optimal design of functionally graded materials using a procedural model and particle
swarm optimization. Computer-Aided Design 44, 4, 300–310.
K RYSL , P., L ALL , S., AND M ARSDEN , J. E. 2001. Dimensional
model reduction in non-linear finite element dynamics of solids
and structures. International Journal for Numerical Methods in
Engineering 51, 4, 479–504.
L I , S., H UANG , J., DE G OES , F., J IN , X., BAO , H., AND D ES BRUN , M. 2014. Space-time editing of elastic motion through
material optimization and reduction. ACM Trans. Graph. 33, 4
(July).
M ANDEL , J. 1993. Balancing domain decomposition. Communications in numerical methods in engineering 9, 3, 233–241.
M ARSDEN , J., AND H UGHES , T. 2012. Mathematical Foundations
of Elasticity. Dover Civil and Mechanical Engineering.
M C A DAMS , A., Z HU , Y., S ELLE , A., E MPEY, M., TAMSTORF,
R., T ERAN , J., AND S IFAKIS , E. 2011. Efficient elasticity
for character skinning with contact and collisions. ACM Trans.
Graph. 30, 4 (July), 37:1–37:12.
NARAIN , R., S AMII , A., AND O’B RIEN , J. F. 2012. Adaptive
anisotropic remeshing for cloth simulation. ACM Trans. Graph.
31, 6 (Nov.), 152:1–152:10.
NARAIN , R., P FAFF , T., AND O’B RIEN , J. F. 2013. Folding
and crumpling adaptive sheets. ACM Trans. Graph. 32, 4 (July),
51:1–51:8.
N ESME , M., K RY, P. G., J E Ř ÁBKOV Á , L., AND FAURE , F. 2009.
Preserving topology and elasticity for embedded deformable models. ACM Trans. Graph. 28, 3 (July), 52:1–52:9.
PATTERSON , T., M ITCHELL , N., AND S IFAKIS , E. 2012. Simulation of complex nonlinear elastic bodies using lattice deformers.
ACM Trans. Graph. 31, 6 (Nov.), 197:1–197:10.
P ERAIRE , J., P EIRO , J., AND M ORGAN , K. 1992. A 3D finite
element multigrid solver for the Euler equations. AIAA paper 92,
0449.
S EILER , M., S PILLMANN , J., AND H ARDERS , M. 2012. Enriching
coarse interactive elastic objects with high-resolution data-driven
deformations. In Proc. ACM SIGGRAPH/Eurographics Symposium on Computer Animation, 9–17.
FARMER , C. 2002. Upscaling: a review. International Journal for
Numerical Methods in Fluids 40, 1-2, 63–78.
S HABANA , A., 1991. Theory of vibration, vol. ii.
FAURE , F., G ILLES , B., B OUSQUET, G., AND PAI , D. K. 2011.
Sparse meshless models of complex deformable solids. ACM
Trans. Graph. 30, 4 (July), 73:1–73:10.
S KOURAS , M., T HOMASZEWSKI , B., C OROS , S., B ICKEL , B.,
AND G ROSS , M. 2013. Computational design of actuated deformable characters. ACM Trans. Graph. 32, 4 (July), 82:1–82:9.
G ILLES , B., B OUSQUET, G., FAURE , F., AND PAI , D. K. 2011.
Frame-based elastic models. ACM Trans. Graph. 30, 2 (Apr.),
15:1–15:12.
S ZAB Ó , B., D ÜSTER , A., AND R ANK , E. 2004. The p-version
of the finite element method. Encyclopedia of computational
mechanics.
G RINSPUN , E., K RYSL , P., AND S CHR ÖDER , P. 2002. Charms: A
simple framework for adaptive simulation. ACM Trans. Graph.
21, 3 (July), 281–290.
W ÄCHTER , A., AND B IEGLER , L. T. 2006. On the implementation
of an interior-point filter line-search algorithm for large-scale
nonlinear programming. Mathematical programming 106, 1.
G UEDES , J., AND K IKUCHI , N. 1990. Preprocessing and postprocessing for materials based on the homogenization method with
adaptive finite element methods. Computer methods in applied
mechanics and engineering 83, 2.
W ICKE , M., R ITCHIE , D., K LINGNER , B. M., B URKE , S.,
S HEWCHUK , J. R., AND O’B RIEN , J. F. 2010. Dynamic local
remeshing for elastoplastic simulation. ACM Trans. Graph. 29, 4
(July), 49:1–49:11.
K AVAN , L., G ERSZEWSKI , D., BARGTEIL , A. W., AND S LOAN ,
P.-P. 2011. Physics-inspired upsampling for cloth simulation in
games. In ACM SIGGRAPH 2011 Papers, ACM, New York, NY,
USA, SIGGRAPH ’11.
X U , H., L I , Y., C HEN , Y., AND BARBI Č , J. 2014. Interactive
material design using model reduction. Tech. rep., University of
Southern California.
K HAREVYCH , L., M ULLEN , P., OWHADI , H., AND D ESBRUN , M.
2009. Numerical coarsening of inhomogeneous elastic materials.
Z HU , Y., S IFAKIS , E., T ERAN , J., AND B RANDT, A. 2010. An
efficient multigrid method for the simulation of high-resolution
elastic solids. ACM Trans. Graph. 29, 2 (Apr.), 16:1–16:18.